Yuandong Tian at AI Frontiers: AI in Games: Achievements and Challenges

AIinGames:AchievementsandChallenges

YuandongTianFacebookAIResearch

GameasaVehicleofAI

Lesssafetyandethical concerns

Fasterthanreal-time

Infinite supplyoffully labeleddata

Controllable and replicable Low costper sample

Complicateddynamicswithsimplerules.

GameasaVehicleofAI

？

Abstractgametoreal-world

Algorithmisslowanddata-inefficient Requirealotofresources.

Hardtobenchmarktheprogress

GameasaVehicleofAI

BetterGames

？




2000s 2010s1990s1970s

GameSpectrum

Good old days 1980s

2000s 2010s1990s1970s

GameSpectrum

Good old days 1980s

ChessGo Poker

2000s 2010s1990s1970s

GameSpectrum

Good old days 1980s

Breakout (1978)Pong (1972)

2000s 2010s1990s1970s

GameSpectrum

Good old days 1980s

Super Mario Bro (1985) Contra (1987)

2000s 2010s1990s1970s

GameSpectrum

Good old days 1980s

Doom (1993) KOF’94 (1994) StarCraft (1998)

2000s 2010s1990s1970s

GameSpectrum

Good old days 1980s

Counter Strike (2000) The Sims 3 (2009)

2000s 2010s1990s1970s

GameSpectrum

Good old days 1980s

GTA V (2013)StarCraft II (2010) Final Fantasy XV (2016)

GameasaVehicleofAI

BetterAlgorithm/System BetterEnvironment

？




Ourwork

DarkForest GoEngine(YuandongTian,YanZhu,ICLR16)

DoomAI(Yuxin Wu,YuandongTian,ICLR17)

ELF:ExtensiveLightweightandFlexibleFramework(YuandongTianetal,arXiv)

BetterAlgorithm/System BetterEnvironment

HowGameAIworks

Evenwithasuper-supercomputer,itisnotpossibletosearchtheentirespace.

HowGameAIworks

Extensivesearch Evaluate

Evenwithasuper-supercomputer,itisnotpossibletosearchtheentirespace.

Consequence

Blackwins

Whitewins

Blackwins

Whitewins

Blackwins

Currentgamesituation

Lufei Ruan vs. Yifan Hou (2010)

HowGameAIworks

Extensivesearch Evaluate Consequence

Blackwins

Whitewins

Blackwins

Whitewins

Blackwins

Howmanyactiondoyouhaveperstep?Checker:a few possiblemovesPoker:a few possiblemovesChess:30-40 possiblemovesGo:100-200 possiblemovesStarCraft:50^100possiblemoves

Alpha-betapruning+Iterativedeepening[MajorChessengine]

Monte-CarloTreeSearch+UCBexploration[MajorGoengine]???

CounterfactualRegretMinimization[Libratus,DeepStack]


HowGameAIworks

Extensivesearch Evaluate Consequence

Blackwins

Whitewins

Blackwins

Whitewins

Blackwins

Howcomplicatedisthegamesituation?Howdeepisthegame?

ChessGoPokerStarCraft

Linearfunctionforsituationevaluation[Stockfish]

DeepValuenetwork[AlphaGo,DeepStack]

Randomgameplaywithsimplerules[Zen,CrazyStone,DarkForest]

Endgamedatabase

Rule-based


HowtomodelPolicy/Valuefunction?

• Manymanualsteps• Conflictingparameters,notscalable.• Needstrongdomainknowledge.

• End-to-Endtraining• Lotsofdata,lesstuning.

• Minimaldomainknowledge.• Amazingperformance

Traditionalapproach DeepLearning

Non-smooth+high-dimensionalSensitivetosituations.OnestonechangesinGoleadstodifferentgame.

Case study: AlphaGo• Computations

• Trainwithmany GPUsandinferencewithTPU.

• Policynetwork• Trainedsupervisedfromhumanreplays.• Self-playnetworkwithRL.

• Highqualityplayout/rolloutpolicy• 2microsecondpermove,24.2%accuracy. ~30%• ThousandsoftimesfasterthanDCNNprediction.

• Valuenetwork• Predictsgameconsequenceforcurrentsituation.• Trainedon30Mself-playgames.

“Mastering the game of Go with deep neural networks and tree search”, Silver et al, Nature 2016

AlphaGo• PolicynetworkSL(trainedwithhumangames)


AlphaGo• FastRollout(2microsecond),~30% accuracy


MonteCarloTreeSearch

2/10

2/10

2/10

1/1

20/30

10/18

9/10

10/12

1/8

22/40

1/1

2/10

2/10

1/1

20/30

10/18

9/10

10/12

1/8

22/402/10

1/1

21/31

11/19

10/11

10/12

1/8

23/41

1/1

(a) (b) (c)

Tree policyDefault policy

Aggregatewinrates,andsearchtowardsthegoodnodes.

PUCT

AlphaGo• ValueNetwork (trained via 30M self-played games)• How data are collected?

Game start

Current state

Sampling SL network(more diverse moves)

Game terminates

Sampling RL network (higher win rate)Uniformsampling


AlphaGo• ValueNetwork (trained via 30M self-played games)


AlphaGo


Ourwork

OurcomputerGoplayer: DarkForest

• DCNNasatreepolicy• Predictnextkmoves(ratherthannextmove)• Trainedon170kKGSdataset/80kGoGoD,57.1% accuracy.• KGS3Dwithoutsearch(0.1spermove)• Release3monthbeforeAlphaGo,<1%GPUs(fromAjaHuang)

Yuandong Tian and Yan Zhu, ICLR 2016

Yan Zhu

Name

Our/enemyliberties

Ko location

Our/enemystones/emptyplace

Our/enemystonehistory

Opponent rank

FeatureusedforDCNN


PureDCNN

WinratebetweenDCNNandopensourceengines.

darkforest:Onlyusetop-1prediction,trainedonKGSdarkfores1:Usetop-3prediction,trainedonGoGoDdarkfores2:darkfores1 withfine-tuning.

MonteCarloTreeSearch

2/10

2/10

2/10

1/1

20/30

10/18

9/10

10/12

1/8

22/40

1/1

2/10

2/10

1/1

20/30

10/18

9/10

10/12

1/8

22/402/10

1/1

21/31

11/19

10/11

10/12

1/8

23/41

1/1

(a) (b) (c)

Tree policyDefault policy

Aggregatewinrates,andsearchtowardsthegoodnodes.

WinratebetweenDCNN+MCTSandopensourceengines.

darkfmcts3:Top-3/5,75krollouts,~12sec/move,KGS5d

94.2%

DCNN + MCTS

• DCNN+MCTS• Usetop3/5movesfromDCNN,75krollouts.• StableKGS5d.Opensource.• 3rd placeonKGSJanuaryTournaments• 2nd placein9th UECComputerGoCompetition(NotthistimeJ)

DarkForest versusKoichiKobayashi(9p)


https://github.com/facebookresearch/darkforestGo

WinRateanalysis(usingDarkForest)(AlphaGo versusLeeSedol)

https://github.com/facebookresearch/ELF/tree/master/go

New version of DarkForest on ELF platform

FirstPersonShooter(FPS)Game

Playthegamefromthe raw image!

Yuxin Wu, Yuandong Tian, ICLR 2017

Yuxin Wu

NetworkStructure

SimpleFrameStacking isveryuseful(ratherthanUsingLSTM)

Actor-Critic Models

Update Policynetwork

UpdateValuenetwork

Reward

Encourageactionsleadingtostateswithhigh-than-expectedvalue.Encouragevaluefunctiontoconvergetothetruecumulativerewards.Keepthediversityofactions

sT

s0

V (sT )

CurriculumTraining

Fromsimpletocomplicated

Curriculum Training

VizDoom AICompetition2016 (Track1)

Rank Bot 1 2 3 4 5 6 7 8 9 10 11 Totalfrags

1 F1 56 62 n/a 54 47 43 47 55 50 48 50 559

2 Arnold 36 34 42 36 36 45 36 39 n/a 33 36 413

3 CLYDE 37 n/a 38 32 37 30 46 42 33 24 44 393

Videos:https://www.youtube.com/watch?v=94EPSjQH38Yhttps://www.youtube.com/watch?v=Qv4esGWOg7w&t=394s

Wewonthefirstplace!

VisualizationofValuefunctionsBest4frames(agentisabouttoshoottheenemy)

Worst4frames(agentmissedtheshootandisoutofammo)

ELF:Extensive,LightweightandFlexibleFrameworkforGameResearch

• Extensive• AnygameswithC++interfacescanbeincorporated.

• Lightweight• Fast.Mini-RTS(40KFPSpercore)• Minimalresourceusage(1GPU+severalCPUs)• Fast training (a couple of hours for a RTS game)

• Flexible• Environment-Actortopology• Parametrized gameenvironments.• ChoiceofdifferentRLmethods.

Yuandong Tian, Qucheng Gong, Wendy Shang, Yuxin Wu, Larry Zitnick (NIPS 2017 Oral)

Arxiv:https://arxiv.org/abs/1707.01067

Larry Zitnick

Qucheng Gong Wendy Shang

Yuxin Wu

https://github.com/facebookresearch/ELF

How RL system works

Game1

GameN

Game2

Consumers (Python)

Actor

Model

Optimizer

Process 1

Process 2

Process N

Replay Buffer

ELFdesign

Plug-and-play; no worry about the concurrency anymore.

Game1

GameN

Daemon(batch

collector)

Producer(GamesinC++)

Game2

History buffer

History buffer

History buffer

Consumers (Python)

Reply

BatchwithHistoryinfo

Actor

Model

Optimizer

PossibleUsage

• GameResearch• Boardgame(Chess,Go,etc)• Real-timeStrategyGame

• Complicated RL algorithms.• Discrete/Continuouscontrol

• Robotics

• DialogandQ&ASystem

Initialization

MainLoop

Training

Self-Play

Multi-Agent

Monte-Carlo Tree Search

1/1

2/10

2/10

1/1

20/30

10/18

9/10

10/12

1/8

22/40

FlexibleEnvironment-Actortopology

(b) Many-to-One (c) One-to-Many

Environment Actor

(a) One-to-OneVanilla A3C BatchA3C, GA3C Self-Play,

Monte-Carlo Tree Search

Environment Actor

Environment Actor

Environment

Environment Actor

Environment

Actor

Environment Actor

Actor

RLPytorch

• ARLplatforminPyTorch• A3Cin30lines.• Interfacingwithdict.

Architecture Hierarchy

ELF

RTS EngineALEGo(DarkForest)

Mini-RTS Capturethe Flag

TowerDefense

An extensive framework that can host many games.

Specific game engines.

EnvironmentsPong Breakout

AminiatureRTSengine

Enemy base

Your base

Your barracks

Worker

Enemy unit

Resource Game Name Descriptions Avg Game Length

Mini-RTS Gather resource and build troops to destroy opponent’s base.

1000-6000 ticks

Capture the Flag Capture the flag and bring it to your own base

1000-4000 ticks

Tower Defense Builds defensive towers to block enemy invasion.

1000-2000 ticks

Fog of War

Simulation Speed

Platform ALE RLE Universe Malmo

FPS 6000 530 60 120

Platform DeepMind Lab VizDoom TorchCraft Mini-RTSFPS 287(C) / 866(G)

6CPU+1GPU7,000 2,000 (Frameskip=50) 40,000

Training AI

Conv ReLUBN

x4

Policy

Value

Game visualization Game internal data(respectingfogofwar)

Location of all workers

Location of all melee tanks

Location of all range tanks

HP portion

UsingInternal Game data and A3C.Rewardisonlyavailableoncethegameisover.

Resource

MiniRTS

Buildingthatcanbuildworkersandcollectresources.

Resourceunitthatcontains1000minerals.

Workerwhocanbuildbarracksandgatherresource.Lowspeedin movement andlowattackdamage.

Buildingthatcanbuildmeleeattackerandrangeattacker.

TankwithhighHP,mediummovementspeed,shortattackrange,highattackdamage.

TankwithlowHP,highmovementspeed,longattackrangeandmediumattackdamage.

Training AI9 discrete actions.

No. Action name Descriptions

1 IDLE Do nothing

2 BUILDWORKER If the base is idle, build a worker

3 BUILDBARRACK Moveaworker(gatheringoridle)toanemptyplaceandbuildabarrack.

4 BUILDMELEEATTACKER Ifwehaveanidlebarrack,buildanmeleeattacker.

5 BUILDRANGEATTACKER Ifwehaveanidlebarrack,builda range attacker.

6 HITANDRUNIfwehaverangeattackers,movetowardsopponentbaseandattack.Takeadvantageoftheirlongattackrangeandhighmovementspeedtohitandrunifenemycounter-attack.

7 ATTACK Allmeleeandrangeattackersattacktheopponent’sbase.

8 ATTACKINRANGE Allmeleeandrangeattackersattackenemiesinsight.

9 ALLDEFEND Alltroopsattackenemytroopsnearthebaseandresource.

Win rate against rule-based AI

Frame skip AI_SIMPLE AI_HIT_AND_RUN

50 68.4(±4.3) 63.6(±7.9)

20 61.4(±5.8) 55.4(±4.7)

10 52.8(±2.4) 51.1(±5.0)

SIMPLE(median)

SIMPLE(mean/std)

HIT_AND_RUN(median)

HIT_AND_RUN(mean/std)

ReLU 52.8 54.7(±4.2) 60.4 57.0(±6.8)

Leaky ReLU 59.8 61.0(±2.6) 60.2 60.3(±3.3)

ReLU + BN 61.0 64.4(±7.4) 55.6 57.5(±6.8)

Leaky ReLU + BN 72.2 68.4(±4.3) 65.5 63.6(±7.9)

Frame skip (how often AI makes decisions)

Network Architecture Conv ReLUBN

Effect of T-steps

Large T is better.

Transfer Learning and Curriculum TrainingAI_SIMPLE AI_HIT_AND_RUN Combined

(50%SIMPLE+50% H&R)

SIMPLE 68.4(±4.3) 26.6(±7.6) 47.5(±5.1)

HIT_AND_RUN 34.6(±13.1) 63.6(±7.9) 49.1(±10.5)

Combined(No curriculum) 49.4(±10.0) 46.0(±15.3) 47.7(±11.0)

Combined 51.8(±10.6) 54.7(±11.2) 53.2(±8.5)

AI_SIMPLE AI_HIT_AND_RUN CAPTURE_THE_FLAG

Withoutcurriculum training 66.0 (±2.4) 54.4 (±15.9) 54.2(±20.0)

Withcurriculum training 68.4 (±4.3) 63.6(±7.9) 59.9(±7.4)

MixtureofSIMPLE_AIandTrainedAI

Trainingtime

99%

Highest win rate against AI_SIMPLE: 80%

Monte Carlo Tree Search

MiniRTS (AI_SIMPLE) MiniRTS (Hit_and_Run)Random 24.2 (±3.9) 25.9 (±0.6)MCTS 73.2 (±0.6) 62.7 (±2.0)

MCTS uses complete information and perfect dynamicsMCTS evaluation is repeated on 1000 games, using 800 rollouts.

Ongoing Work• One framework for different games.

• DarkForest remastered: https://github.com/facebookresearch/ELF/tree/master/go

• Richergamescenarios for MiniRTS.• Multiplebases(Expand?Rush?Defending?)• Morecomplicatedunits.• Provide a LUA interface for easier modification of the game.

• Realisticactionspace• Onecommand perunit

• Model-basedReinforcementLearning• MCTS with perfect information and perfect dynamics also achieves ~70% winrate

• Self-Play(TrainedAIversusTrainedAI)

Thanks!

Yuandong Tian at AI Frontiers: AI in Games: Achievements and Challenges

Data & Analytics